Bench
What actually measured well.
One dot per measured configuration: accuracy across, joules per correct answer up. Down and to the right is better. Every dot is a run submitted through the same open benchmark harness this project ships, not a spec-sheet estimate; hover one for its numbers, or open the table.
Reading the chart. The amber dots joined by a stepped line are the Pareto front: nothing measured beats them on both axes at once. A hollow dot is a quantized checkpoint. TTFT vs accuracy and TTFT vs tok/s move energy into the dot colour so responsiveness can sit on an axis.
Why one benchmark at a time. Joules per correct answer is only comparable within a task: the same model needs a different amount of work for an mmlu answer and a gsm8k one. Pooling tasks would draw a chart whose real axis is "which benchmark".
Chance floor. On a four-choice task a model that guesses still books about 25 correct answers per 100 for almost no energy, so its J/correct is tiny and it sorts to the top. A cell only ranks when its accuracy interval's lower bound clears the floor. Unticked, sub-floor cells are drawn with a red ring and kept off the front.
Model source and size are read from the model's name (its organisation prefix and the parameter count written into it, e.g. 8B); a name that says neither shows under "other" and no size bucket. Tier is the measurement tier: A has GPU counter, power samples and a wall plug; C is GPU-only.
Seed rows are our own reference matrix; community rows appear from the first contributed run, and each one's tooltip says how many distinct machines back it.